A method and apparatus for generating multi-source fusion features for transformer condition assessment

Abstract

This invention discloses a method and apparatus for generating multi-source fusion features for transformer condition assessment. The method includes acquiring condition data from various data acquisition components associated with the transformer according to a preset acquisition cycle; reducing the dimensionality of each condition data based on a constructed dual sparse dictionary model to obtain dimensionality-reduced data; weighting and filling the dimensionality-reduced data according to various filling methods to obtain filled data; and cleaning the filled data based on the reconstruction error between the filled data and the condition data to generate multi-source fusion features, thereby timely and effectively preprocessing transformer data and effectively improving the accuracy and reliability of transformer assessment.

Description

Technical Field

[0001] This invention relates to the field of feature processing technology, and in particular to a method and apparatus for generating multi-source fusion features for transformer condition assessment. Background Technology

[0002] In today's power systems, transformers are critical equipment, and their operating status has a vital impact on the normal operation of the power system. However, existing transformer evaluation systems have some significant shortcomings.

[0003] Current transformer condition assessment systems typically only consider electrical test data and oil chromatography data as input for characteristic parameters. Furthermore, during data preprocessing, only one data imputation method is generally used for missing data, and rough set theory and deep belief networks are typically employed for sparse dimensionality reduction. For data cleaning, an autoencoder is used.

[0004] The above-mentioned single data source selection has significant limitations and cannot fully reflect the actual state of the transformer under overall operating conditions. Considering only a single filling method cannot fill from the perspective of the entire data level, which may lead to overfitting during training. Existing dimensionality reduction methods do not consider the impact of time-series historical data. Data cleaning using autoencoders has high computational resource requirements and makes it difficult to perform effective data preprocessing of transformer data in a timely manner, which can easily lead to a decrease in the accuracy and reliability of transformer evaluation. Summary of the Invention

[0005] This invention provides a multi-source fusion feature generation method and apparatus for transformer condition assessment. It solves the technical problems in the existing transformer condition assessment process, such as the great limitations of selecting a single data source, which cannot fully reflect the actual state of the transformer under the overall operating conditions; considering only a single filling method, which cannot fill from the perspective of the entire data level, which may lead to training overfitting; existing dimensionality reduction methods do not consider the influence of time-series historical data; data cleaning using autoencoders has high requirements for computing resources, and it is difficult to perform effective data preprocessing of transformer data in a timely manner, which can easily lead to a decrease in the accuracy and reliability of transformer assessment.

[0006] This invention provides a multi-source fusion feature generation method for transformer condition assessment, comprising:

[0007] According to the preset acquisition cycle, status data is acquired from various data acquisition components associated with the transformer.

[0008] The dimensionality of each state data is reduced based on the constructed dual sparse dictionary model to obtain dimensionality-reduced data;

[0009] The dimensionality-reduced data is then weighted and filled using multiple filling methods to obtain filled data.

[0010] Based on the reconstruction error between the filled data and the state data, the filled data is cleaned to generate multi-source fusion features.

[0011] Optionally, the step of acquiring status data from various data acquisition components associated with the transformer according to a preset acquisition cycle includes:

[0012] The transformer is connected to multiple data acquisition components to collect feature data in real time; the feature data includes sound signals, light signals, ultrasonic signals and vibration signals.

[0013] When the acquisition time meets the preset acquisition cycle, the sound signal, the light signal, the ultrasonic signal and the vibration signal are acquired from each of the data acquisition components respectively;

[0014] The sound signal, light signal, ultrasonic signal, and vibration signal are sorted according to the acquisition time to construct state data in sequence form.

[0015] Optionally, the dimensionality reduction of each state data based on the constructed dual sparse dictionary model is performed to obtain dimensionality-reduced data, including:

[0016] A dual sparse dictionary model is constructed by superimposing a learned dictionary model on a fixed dictionary model as the base layer.

[0017] Project each of the state data onto the dual sparse dictionary model;

[0018] The orthogonal matching pursuit algorithm is used to solve the sparse representation of each state data under the dual sparse dictionary model to obtain dimensionality-reduced data.

[0019] The dual sparse dictionary model is as follows:

[0020]

[0021] in, It is a coefficient matrix The Middle The coefficient column vector corresponding to each dictionary element State data in a learned dictionary model The sparse representation coefficient matrix below, It is a learning dictionary model The coefficient matrix after sparse representation, It is a fixed dictionary model with each dictionary element predefined. It is the i-th sample in the state data. It is a sparse coefficient vector. It is a sparse representation of state data, i.e., dimensionality-reduced data. This represents the number of non-zero elements in the sparse coefficient vector.

[0022] Optionally, the step of weighting and filling the dimensionality-reduced data according to multiple filling methods to obtain filled data includes:

[0023] Calculate the simulated fill values ​​corresponding to the dimensionality reduction data according to various fill methods;

[0024] The error coefficients corresponding to various filling methods are calculated by using each of the simulated filling values ​​and the known values ​​in the dimensionality reduction data, respectively.

[0025] The fill weights for each of the aforementioned fill methods are set according to the respective error coefficients;

[0026] The actual fill value is obtained by weighting the fill values ​​of each fill method according to the fill weight.

[0027] The missing values ​​in the dimensionality-reduced data are filled with the actual filling values ​​to obtain the filled data.

[0028] Optionally, the step of cleaning the fill data and generating multi-source fusion features based on the reconstruction error between the fill data and the state data includes:

[0029] Calculate the loss error between the filling data and the state data during the acquisition period;

[0030] If none of the loss errors exceed the preset loss threshold, the filled data is determined to be a lossless feature;

[0031] If the loss error exceeds a preset loss threshold, then the error duration is recorded.

[0032] If the error duration does not exceed the preset duration threshold, the filling data is determined to be a fluctuation feature;

[0033] If the error duration exceeds a preset duration threshold, a preset stacked denoising autoencoder model is invoked to reconstruct the filled data and obtain reconstructed features.

[0034] Multi-source fusion features are generated using the lossless features, the fluctuation features, and / or the reconstructed features.

[0035] The present invention also provides a multi-source fusion feature generation device for transformer condition assessment, comprising:

[0036] The status data acquisition module is used to acquire status data from various data acquisition components associated with the transformer according to a preset acquisition cycle.

[0037] The data dimensionality reduction module is used to reduce the dimensionality of each state data based on the constructed dual sparse dictionary model to obtain dimensionality-reduced data;

[0038] The data filling module is used to perform weighted filling on the dimensionality-reduced data according to multiple filling methods to obtain filled data;

[0039] The data cleaning module is used to clean the fill data based on the reconstruction error between the fill data and the state data, and generate multi-source fusion features.

[0040] Optionally, the status data acquisition module is specifically used for:

[0041] The transformer is connected to multiple data acquisition components to collect feature data in real time; the feature data includes sound signals, light signals, ultrasonic signals and vibration signals.

[0042] When the acquisition time meets the preset acquisition cycle, the sound signal, the light signal, the ultrasonic signal and the vibration signal are acquired from each of the data acquisition components respectively;

[0043] The sound signal, light signal, ultrasonic signal, and vibration signal are sorted according to the acquisition time to construct state data in sequence form.

[0044] Optionally, the data dimensionality reduction module is specifically used for:

[0045] A dual sparse dictionary model is constructed by superimposing a learned dictionary model on a fixed dictionary model as the base layer.

[0046] Project each of the state data onto the dual sparse dictionary model;

[0047] The orthogonal matching pursuit algorithm is used to solve the sparse representation of each state data under the dual sparse dictionary model to obtain dimensionality-reduced data.

[0048] The dual sparse dictionary model is as follows:

[0049]

[0050] in, It is a coefficient matrix The Middle The coefficient column vector corresponding to each dictionary element State data in a learned dictionary model The sparse representation coefficient matrix below, It is a learning dictionary model The coefficient matrix after sparse representation, It is a fixed dictionary model with each dictionary element predefined. It is the i-th sample in the state data. It is a sparse coefficient vector. It is a sparse representation of state data, i.e., dimensionality-reduced data. This represents the number of non-zero elements in the sparse coefficient vector.

[0051] Optionally, the data filling module is specifically used for:

[0052] Calculate the simulated fill values ​​corresponding to the dimensionality reduction data according to various fill methods;

[0053] The error coefficients corresponding to various filling methods are calculated by using each of the simulated filling values ​​and the known values ​​in the dimensionality reduction data, respectively.

[0054] The fill weights for each of the aforementioned fill methods are set according to the respective error coefficients;

[0055] The actual fill value is obtained by weighting the fill values ​​of each fill method according to the fill weight.

[0056] The missing values ​​in the dimensionality-reduced data are filled with the actual filling values ​​to obtain the filled data.

[0057] Optionally, the data cleaning module is specifically used for:

[0058] Calculate the loss error between the filling data and the state data during the acquisition period;

[0059] If none of the loss errors exceed the preset loss threshold, the filled data is determined to be a lossless feature;

[0060] If the loss error exceeds a preset loss threshold, then the error duration is recorded.

[0061] If the error duration does not exceed the preset duration threshold, the filling data is determined to be a fluctuation feature;

[0062] If the error duration exceeds a preset duration threshold, a preset stacked denoising autoencoder model is invoked to reconstruct the filled data and obtain reconstructed features.

[0063] Multi-source fusion features are generated using the lossless features, the fluctuation features, and / or the reconstructed features.

[0064] As can be seen from the above technical solutions, the present invention has the following advantages:

[0065] This invention acquires state data from various data acquisition components associated with the transformer according to a preset acquisition cycle; performs dimensionality reduction on each state data based on a constructed dual sparse dictionary model to obtain dimensionality-reduced data; performs weighted filling on the dimensionality-reduced data according to various filling methods to obtain filled data; and cleans the filled data based on the reconstruction error between the filled data and the state data to generate multi-source fusion features, thereby timely and effectively preprocessing the transformer data and effectively improving the accuracy and reliability of transformer evaluation. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 A flowchart illustrating the steps of a multi-source fusion feature generation method for transformer condition assessment provided in this embodiment of the invention;

[0068] Figure 2 This is a structural block diagram of a multi-source fusion feature generation device for transformer condition assessment provided in an embodiment of the present invention. Detailed Implementation

[0069] This invention provides a multi-source fusion feature generation method and apparatus for transformer condition assessment. It addresses the limitations of existing transformer condition assessment methods, such as the significant limitations of selecting a single data source (failing to comprehensively reflect the actual state of the transformer under overall operating conditions), the inability to fill data from a single perspective (leading to overfitting), the lack of consideration for the impact of historical time-series data in existing dimensionality reduction methods, and the high computational resource requirements of autoencoder-based data cleaning, which hinders timely and effective data preprocessing of transformer data and reduces the accuracy and reliability of transformer assessment.

[0070] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0071] Please see Figure 1 , Figure 1 The flowchart illustrates the steps of a multi-source fusion feature generation method for transformer condition assessment provided in this embodiment of the invention.

[0072] This invention provides a multi-source fusion feature generation method for transformer condition assessment, comprising:

[0073] Step 101: Acquire status data from various data acquisition components associated with the transformer according to the preset acquisition cycle;

[0074] A data acquisition component refers to a modular component used to acquire various signal data. It can be powered independently or by the processor power supply in this embodiment, and this embodiment does not limit this. Its types may include, but are not limited to, sound signal acquisition components such as microphones and sound cards, optical signal acquisition components such as photodiodes and image sensors, ultrasonic signal acquisition components such as ultrasonic transducers and signal conditioning circuits, and vibration signal acquisition components such as accelerometers and data acquisition cards.

[0075] Status data refers to various serialized data used to assess the operating status of a transformer, including but not limited to serialized data of sound signals, light signals, ultrasonic signals, and vibration signals.

[0076] In this embodiment, the method can be applied to a central processing unit with communication and data processing functions. In order to obtain state data in the form of time series, various characteristic data of the transformer can be collected in real time through data acquisition components. The processor obtains state data from various data acquisition components associated with the transformer according to a preset acquisition cycle as the data basis.

[0077] In one example of this application, step 101 may include the following sub-steps:

[0078] The transformer is connected to various data acquisition components to collect feature data in real time; the feature data includes sound signals, light signals, ultrasonic signals and vibration signals.

[0079] When the acquisition time meets the preset acquisition cycle, sound signals, light signals, ultrasonic signals and vibration signals are acquired from each data acquisition component respectively;

[0080] The sound signals, light signals, ultrasonic signals, and vibration signals were sorted according to the acquisition time, and state data in sequence form were constructed respectively.

[0081] Existing transformer condition assessments mostly consider only electrical test data and oil chromatography data. In this embodiment, to achieve real-time monitoring and multi-source fusion of the transformer, multiple data acquisition components associated with the transformer can be simultaneously invoked during transformer startup to collect multiple characteristic data of the transformer in real time. When the acquisition time meets the preset acquisition cycle, or when multi-source fusion features need to be generated, such as when an external request or trigger signal is received, the processor acquires sound signals, light signals, ultrasonic signals, and vibration signals from each data acquisition component as the data basis. Since the above signals are acquired in chronological order, they can be sorted according to the acquisition time to obtain state data in sequence form.

[0082] It should be noted that the normal operating noise of electrical equipment such as transformers is a low-frequency audible sound, while the energy of the discharge sound signal is mostly concentrated in the high-frequency range. The normal operating sound of a transformer has a relatively uniform energy distribution in the 0-6kHz range, but the main energy is concentrated below 1kHz. In contrast, the high-energy part of the spark discharge sound is distributed around 10kHz, showing a significant difference in sound characteristics. The light signal generated during corona discharge has both continuous and line spectra, with wavelengths mainly distributed between 280-400nm, belonging to the ultraviolet band, and a small portion of wavelengths distributed between 230-280nm. The energy of the corona current can be reflected by the magnitude of the ultraviolet radiation power, so the magnitude of the ultraviolet radiation power can be used as a basis for assessing the strength of partial discharge. When partial discharge occurs, because the bubble is charged, it will be squeezed by the electric field force, thereby causing changes in the local volume of the bubble. Changes in bubble volume create compression-sparse waves around the transformer, which are ultrasonic waves. The process of partial discharge generating ultrasonic waves is actually a process of electro-mechanical-sound conversion. The amplitude of the core vibration acceleration is proportional to the square of the voltage, and its vibration frequency is twice the voltage frequency. Since the power frequency is 50Hz, the core vibration acceleration is 100Hz. The deformation of the core will intensify its vibration intensity, which will make the amplitude, frequency, and trend of the vibration signal different from the core vibration signal during normal operation. Operators can extract its characteristic values ​​by collecting the vibration signal and distinguish it from the normal vibration signal, thereby diagnosing the transformer core vibration.

[0083] Step 102: Based on the constructed dual sparse dictionary model, the dimensionality of each state data is reduced to obtain dimensionality-reduced data;

[0084] The dual-sparse dictionary model combines a fixed dictionary with a learned dictionary. This novel parameter structure results in a concise and flexible dictionary representation, effectively reducing dictionary training time and enhancing adaptability and effectiveness. In this embodiment, the dual-sparse dictionary model uses a sparse base dictionary as the fixed dictionary and an overcomplete dictionary as the learned dictionary. A sparse base dictionary is a dictionary with specific structure and properties, designed to represent a given signal with as few non-zero coefficients as possible. It typically possesses certain mathematical structures, such as orthogonality and tight-frame properties. For example, Discrete Cosine Transform (DCT) dictionaries and Discrete Wavelet Transform (DWT) dictionaries are common sparse base dictionaries. An overcomplete dictionary refers to a dictionary whose number of elements far exceeds the dimension of the signal; that is, the dictionary is overcomplete.

[0085] In this embodiment, after acquiring the state data in sequence form, due to the large amount of data collected in real time, it is necessary to reduce the dimensionality of the state data to improve the efficiency of subsequent data processing. Specifically, after constructing a dual sparse dictionary model, the dimensionality of each state data is reduced using the dual sparse dictionary model, thereby obtaining the sparse representation of each type of state under the dual sparse dictionary model as the dimensionality-reduced data.

[0086] In one example of this application, step 102 may include the following sub-steps:

[0087] A dual sparse dictionary model is constructed by superimposing a learned dictionary model on a fixed dictionary model as the base layer.

[0088] Project each state data onto the dual sparse dictionary model;

[0089] The orthogonal matching pursuit algorithm is used to solve the sparse representation of each state data under the dual sparse dictionary model, and the dimensionality-reduced data is obtained.

[0090] The double sparse dictionary model is as follows:

[0091]

[0092] in, It is a coefficient matrix The Middle The coefficient column vector corresponding to each dictionary element State data in a learned dictionary model The sparse representation coefficient matrix below, It is a learning dictionary model The coefficient matrix after sparse representation, It is a fixed dictionary model with each dictionary element predefined. It is the i-th sample in the state data. It is a sparse coefficient vector. It is a sparse representation of state data, i.e., dimensionality-reduced data. This represents the number of non-zero elements in the sparse coefficient vector.

[0093] In this embodiment, a fixed dictionary and a learned dictionary can be used simultaneously in signal representation. The final sparse representation is obtained through linear combination or weighted summation. The weight allocation strategy for the fixed and learned dictionaries in the superposition is then determined, and can be adjusted based on prior knowledge or through experimentation. Joint optimization is performed to simultaneously adjust the parameters of the fixed and learned dictionaries, enabling the dual-sparse dictionary model to achieve optimal performance on a specific task. An alternating optimization method can be used, optimizing the fixed and learned dictionaries separately and then alternating until convergence. During the optimization process, appropriate regularization techniques are considered to prevent overfitting. After constructing the dual-sparse dictionary model, the state data is projected onto it. The state data first undergoes feature extraction using the fixed dictionary to extract general prior features, and then the learned dictionary model adaptively extracts features based on the specific type of signal to obtain more targeted features.

[0094] In the specific implementation, after projecting each data point onto the dual sparse dictionary model, the orthogonal matching pursuit algorithm can be used to solve the sparse representation of each state data under the dual sparse dictionary model, thus obtaining the dimensionality-reduced data.

[0095] It should be noted that the Orthogonal Matching Pursuit (OMP) algorithm is used to find a sparse representation of a signal under an overcomplete dictionary. Its basic idea is to select an atom from the dictionary that best matches the current residual each time, then project the signal onto that atom, update the residual, and repeat this process until a certain stopping condition is met.

[0096] In this embodiment, the data dimensionality reduction in data preprocessing differs from existing rough sets and deep belief networks. Simple data dimensionality reduction, which only considers data, cannot take into account time-series historical data and has poor processing effect on large amounts of data. Therefore, a dual sparse dictionary is used to perform feature sparse dimensionality reduction on a large amount of online time-series data of transformers.

[0097] Step 103: Perform weighted filling on the dimensionality-reduced data according to multiple filling methods to obtain the filled data;

[0098] In this embodiment, to address the different signal types in different dimensionality-reduced data, after dimensionality reduction, due to the sparse representation of the data, multiple missing data may appear. To improve the accuracy of subsequent data processing as much as possible, the error coefficients of various filling methods can be weighted to combine the advantages of different filling methods to perform weighted filling of the dimensionality-reduced data, thereby obtaining filled data.

[0099] In one example of this application, step 103 may include the following sub-steps:

[0100] Calculate the simulated fill values ​​corresponding to the dimensionality reduction data according to various fill methods;

[0101] The error coefficients for various filling methods are calculated by using each simulated filling value and the known values ​​in the dimensionality-reduced data respectively.

[0102] Set the fill weights for each fill method according to the error coefficients;

[0103] The actual fill value is obtained by weighting the fill values ​​of each fill method according to the fill weight.

[0104] The missing values ​​in the dimensionality-reduced data are filled with the actual fill values ​​to obtain the filled data.

[0105] In this embodiment, simulated fill values ​​corresponding to the dimensionality-reduced data are calculated according to various fill methods. Error coefficients for each fill method are calculated using both the simulated fill value and known values ​​from the dimensionality-reduced data. These error coefficients can be calculated in various ways, such as by calculating the mean absolute error or mean squared error between the filled value and the true value. After calculating the error coefficients, fill weights for each fill method are assigned according to their magnitude; the smaller the error coefficient, the greater the weight, and the sum of the weights is 1. Fill weights can be assigned by calculating the proportion of the error coefficients and taking the reciprocal.

[0106] After setting the imputation weights, the imputation values ​​for each imputation method are weighted and averaged according to these weights to calculate the actual imputation value at each position. Then, the missing values ​​in the dimensionality-reduced data are imputed using these actual imputation values ​​to generate the imputed data.

[0107] It should be noted that the filling methods can include, but are not limited to, KNN filling, mean filling, zero filling, median filling, forward filling, and backward filling.

[0108] Step 104: Based on the reconstruction error between the filled data and the state data, perform data cleaning on the filled data to generate multi-source fusion features.

[0109] In this embodiment, to reduce the resource consumption of data cleaning, after generating the filler data, the loss error and error duration between the filler data and the state data are calculated to obtain the reconstruction error between the filler data and the state data. The filler data is then cleaned according to the reconstruction error to generate multi-source fusion features.

[0110] In one example of this application, step 104 may include the following sub-steps:

[0111] Calculate the loss error between the filler data and the status data during the acquisition period;

[0112] If the loss error does not exceed the preset loss threshold, the filled data will be determined as a lossless feature;

[0113] If the loss error exceeds the preset loss threshold, the error duration will be recorded.

[0114] If the error duration does not exceed the preset duration threshold, the filled data will be identified as a fluctuation characteristic.

[0115] If the error duration exceeds the preset duration threshold, the preset stacked denoising autoencoder model is invoked to reconstruct the data from the filled data and obtain the reconstructed features.

[0116] Multi-source fusion features are generated by employing lossless features, fluctuation features, and / or reconstructed features.

[0117] The Stacked Denoising Autoencoder (SDAE) model is composed of multiple stacked denoising autoencoders (DAEs). Each DAE contains an encoder and a decoder. The encoder maps the input data to a low-dimensional latent representation, while the decoder attempts to reconstruct the original input data from this latent representation. By stacking multiple DAEs, higher-level feature representations can be learned progressively.

[0118] In this embodiment, the loss error between the fill data and the status data during the acquisition period can be calculated using the following formula:

[0119]

[0120] in, To compensate for the loss error between the fill data and the status data during the acquisition period, This represents a loss function based on a specific function H, whose inputs are state data x and padding data z. The entropy function is used to measure uncertainty. To perform a normalization transformation on the variable x, Similarly.

[0121] In this embodiment, the loss threshold is Thd. Data with a reconstruction error Re ≤ Thd is considered lossless. If Re > Thd and Et ≤ Tw for a portion of the data, it indicates that these monitoring parameters are simultaneously affected, and transformer anomalies or faults cannot be ruled out. If Re > Thd and Et > Tw for a portion of the data, the state data represents the normal operating state of the transformer, and "dirty" data represents missing values. If Re > Thd and Et > Tw for a portion of the data, the state data represents the abnormal operating state of the transformer. The abnormal data defined in the above steps uses the reconstructed representation of "dirty" values ​​from the SDAE model as repair values. Lossless data, along with the repair values ​​for singularities and missing values, constitute valid information for normal state data.

[0122] In practical implementation, since current data cleaning using autoencoders results in high training resource consumption and poor training performance, this embodiment proposes inputting padding data with error durations exceeding a preset duration threshold into a pre-trained SDAE model. The SDAE model processes the input data according to its trained structure and parameters. First, the underlying denoising autoencoder (DAE) encodes the input data, mapping it to a low-dimensional latent representation space. During this process, the model attempts to remove noise and anomalies from the data. Then, the decoder gradually reconstructs the data from the low-dimensional latent representation. Because the SDAE model learns the feature patterns of normal data during training, it reconstructs a data representation close to normal based on the characteristics of the input "dirty" data, thus obtaining reconstructed data.

[0123] It should be noted that the error duration is Et, and Tw is the maximum allowable time period or duration of error. Within this time range, even if the system exhibits some error or deviation, it is considered acceptable and does not require immediate action.

[0124] In this embodiment, state data is acquired from various data acquisition components associated with the transformer according to a preset acquisition cycle. Dimensionality reduction is performed on each state data based on a constructed dual-sparse dictionary model to obtain dimensionality-reduced data. Weighted filling is then performed on the dimensionality-reduced data using various filling methods to obtain filled data. Based on the reconstruction error between the filled data and the state data, the filled data is cleaned to generate multi-source fusion features, thereby enabling timely and effective data preprocessing of the transformer data and significantly improving the accuracy and reliability of transformer evaluation. For the construction of feature parameters in the transformer state evaluation system, a multi-source data feature parameter construction based on the fusion of non-electrical quantity data acquisition is proposed. For data loss, data noise, and sparse dimensionality reduction in data preprocessing, a weighted preprocessing method combining multiple filling methods, a stacked noise-reducing autoencoder data cleaning method, and a dual-sparse dictionary model multi-source data signal compression sensing method are proposed respectively, realizing a multi-system fusion method for overall data preprocessing.

[0125] Please see Figure 2 , Figure 2 A structural block diagram of a multi-source fusion feature generation device for transformer condition assessment is shown in an embodiment of this application.

[0126] This invention provides a multi-source fusion feature generation device for transformer condition assessment, comprising:

[0127] The status data acquisition module 201 is used to acquire status data from various data acquisition components associated with the transformer according to a preset acquisition cycle.

[0128] The data dimensionality reduction module 202 is used to reduce the dimensionality of each state data based on the constructed dual sparse dictionary model to obtain dimensionality-reduced data;

[0129] Data filling module 203 is used to perform weighted filling on dimensionality-reduced data according to multiple filling methods to obtain filled data;

[0130] The data cleaning module 204 is used to clean the fill data based on the reconstruction error between the fill data and the state data, and generate multi-source fusion features.

[0131] Optionally, the status data acquisition module 201 is specifically used for:

[0132] The transformer is connected to various data acquisition components to collect feature data in real time; the feature data includes sound signals, light signals, ultrasonic signals and vibration signals.

[0133] When the acquisition time meets the preset acquisition cycle, sound signals, light signals, ultrasonic signals and vibration signals are acquired from each data acquisition component respectively;

[0134] The sound signals, light signals, ultrasonic signals, and vibration signals were sorted according to the acquisition time, and state data in sequence form were constructed respectively.

[0135] Optionally, the data dimensionality reduction module 202 is specifically used for:

[0136] A dual sparse dictionary model is constructed by superimposing a learned dictionary model on a fixed dictionary model as the base layer.

[0137] Project each state data onto the dual sparse dictionary model;

[0138] The orthogonal matching pursuit algorithm is used to solve the sparse representation of each state data under the dual sparse dictionary model, and the dimensionality-reduced data is obtained.

[0139] The double sparse dictionary model is as follows:

[0140]

[0141] in, It is a coefficient matrix The Middle The coefficient column vector corresponding to each dictionary element State data in a learned dictionary model The sparse representation coefficient matrix below, It is a learning dictionary model The coefficient matrix after sparse representation, It is a fixed dictionary model with each dictionary element predefined. It is the i-th sample in the state data. It is a sparse coefficient vector. It is a sparse representation of state data, i.e., dimensionality-reduced data. This represents the number of non-zero elements in the sparse coefficient vector.

[0142] Optionally, the data filling module 203 is specifically used for:

[0143] Calculate the simulated fill values ​​corresponding to the dimensionality reduction data according to various fill methods;

[0144] The error coefficients for various filling methods are calculated by using each simulated filling value and the known values ​​in the dimensionality-reduced data respectively.

[0145] Set the fill weights for each fill method according to the error coefficients;

[0146] The actual fill value is obtained by weighting the fill values ​​of each fill method according to the fill weight.

[0147] The missing values ​​in the dimensionality-reduced data are filled with the actual fill values ​​to obtain the filled data.

[0148] Optionally, the data cleaning module 204 is specifically used for:

[0149] Calculate the loss error between the filler data and the status data during the acquisition period;

[0150] If the loss error does not exceed the preset loss threshold, the filled data will be determined as a lossless feature;

[0151] If the loss error exceeds the preset loss threshold, the error duration will be recorded.

[0152] If the error duration does not exceed the preset duration threshold, the filled data will be identified as a fluctuation characteristic.

[0153] If the error duration exceeds the preset duration threshold, the preset stacked denoising autoencoder model is invoked to reconstruct the data from the filled data and obtain the reconstructed features.

[0154] Multi-source fusion features are generated by employing lossless features, fluctuation features, and / or reconstructed features.

[0155] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device and module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0156] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0157] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating multi-source fusion features for transformer condition assessment, characterized in that, include: According to the preset acquisition cycle, status data is acquired from various data acquisition components associated with the transformer. The dimensionality of each state data is reduced based on the constructed dual sparse dictionary model to obtain dimensionality-reduced data; The dimensionality-reduced data is then weighted and filled using multiple filling methods to obtain filled data. Based on the reconstruction error between the filled data and the state data, the filled data is cleaned to generate multi-source fusion features; The step of cleaning the filled data and generating multi-source fusion features based on the reconstruction error between the filled data and the state data includes: Calculate the loss error between the filling data and the state data during the acquisition period; If none of the loss errors exceed the preset loss threshold, the filled data is determined to be a lossless feature; If the loss error exceeds a preset loss threshold, then the error duration is recorded. If the error duration does not exceed the preset duration threshold, the filling data is determined to be a fluctuation feature; If the error duration exceeds a preset duration threshold, a preset stacked denoising autoencoder model is invoked to reconstruct the filled data and obtain reconstructed features. Multi-source fusion features are generated using the lossless features, the fluctuation features, and / or the reconstructed features.

2. The method according to claim 1, characterized in that, The process of acquiring status data from various data acquisition components associated with the transformer according to a preset acquisition cycle includes: The transformer is connected to multiple data acquisition components to collect feature data in real time; the feature data includes sound signals, light signals, ultrasonic signals and vibration signals. When the acquisition time meets the preset acquisition cycle, the sound signal, the light signal, the ultrasonic signal and the vibration signal are acquired from each of the data acquisition components respectively; The sound signal, light signal, ultrasonic signal, and vibration signal are sorted according to the acquisition time to construct state data in sequence form.

3. The method according to claim 1, characterized in that, The constructed dual-sparse dictionary model is used to reduce the dimensionality of each state data to obtain dimensionality-reduced data, including: A dual sparse dictionary model is constructed by superimposing a learned dictionary model on a fixed dictionary model as the base layer. Project each of the state data onto the dual sparse dictionary model; The orthogonal matching pursuit algorithm is used to solve the sparse representation of each state data under the dual sparse dictionary model to obtain dimensionality-reduced data. The dual sparse dictionary model is as follows: ； in, It is a coefficient matrix The Middle The coefficient column vector corresponding to each dictionary element State data in a learned dictionary model The sparse representation coefficient matrix below, It is a learning dictionary model The coefficient matrix after sparse representation, It is a fixed dictionary model with each dictionary element predefined. It is the i-th sample in the state data. It is a sparse coefficient vector. It is a sparse representation of state data, i.e., dimensionality-reduced data. This represents the number of non-zero elements in the sparse coefficient vector.

4. The method according to claim 1, characterized in that, The step of weighting and filling the dimensionality-reduced data according to multiple filling methods to obtain filled data includes: Calculate the simulated fill values ​​corresponding to the dimensionality reduction data according to various fill methods; The error coefficients corresponding to various filling methods are calculated by using each of the simulated filling values ​​and the known values ​​in the dimensionality reduction data, respectively. The fill weights for each of the aforementioned fill methods are set according to the respective error coefficients; The actual fill value is obtained by weighting the fill values ​​of each fill method according to the fill weight. The missing values ​​in the dimensionality-reduced data are filled with the actual filling values ​​to obtain the filled data.

5. A multi-source fusion feature generation device for transformer condition assessment, characterized in that, include: The status data acquisition module is used to acquire status data from various data acquisition components associated with the transformer according to a preset acquisition cycle. The data dimensionality reduction module is used to reduce the dimensionality of each state data based on the constructed dual sparse dictionary model to obtain dimensionality-reduced data; The data filling module is used to perform weighted filling on the dimensionality-reduced data according to multiple filling methods to obtain filled data; The data cleaning module is used to clean the fill data based on the reconstruction error between the fill data and the state data, and generate multi-source fusion features. The data cleaning module is specifically used for: Calculate the loss error between the filling data and the state data during the acquisition period; If none of the loss errors exceed the preset loss threshold, the filled data is determined to be a lossless feature; If the loss error exceeds a preset loss threshold, then the error duration is recorded. If the error duration does not exceed the preset duration threshold, the filling data is determined to be a fluctuation feature; If the error duration exceeds a preset duration threshold, a preset stacked denoising autoencoder model is invoked to reconstruct the filled data and obtain reconstructed features. Multi-source fusion features are generated using the lossless features, the fluctuation features, and / or the reconstructed features.

6. The apparatus according to claim 5, characterized in that, The status data acquisition module is specifically used for: The transformer is connected to multiple data acquisition components to collect feature data in real time; the feature data includes sound signals, light signals, ultrasonic signals and vibration signals. When the acquisition time meets the preset acquisition cycle, the sound signal, the light signal, the ultrasonic signal and the vibration signal are acquired from each of the data acquisition components respectively; The sound signal, light signal, ultrasonic signal, and vibration signal are sorted according to the acquisition time to construct state data in sequence form.

7. The apparatus according to claim 5, characterized in that, The data dimensionality reduction module is specifically used for: A dual sparse dictionary model is constructed by superimposing a learned dictionary model on a fixed dictionary model as the base layer. Project each of the state data onto the dual sparse dictionary model; The orthogonal matching pursuit algorithm is used to solve the sparse representation of each state data under the dual sparse dictionary model to obtain dimensionality-reduced data. The dual sparse dictionary model is as follows: ； in, It is a coefficient matrix The Middle The coefficient column vector corresponding to each dictionary element State data in a learned dictionary model The sparse representation coefficient matrix below, It is a learning dictionary model The coefficient matrix after sparse representation, It is a fixed dictionary model with each dictionary element predefined. It is the i-th sample in the state data. It is a sparse coefficient vector. It is a sparse representation of state data, i.e., dimensionality-reduced data. This represents the number of non-zero elements in the sparse coefficient vector.

8. The apparatus according to claim 5, characterized in that, The data filling module is specifically used for: Calculate the simulated fill values ​​corresponding to the dimensionality reduction data according to various fill methods; The error coefficients corresponding to various filling methods are calculated by using each of the simulated filling values ​​and the known values ​​in the dimensionality reduction data, respectively. The fill weights for each of the aforementioned fill methods are set according to the respective error coefficients; The actual fill value is obtained by weighting the fill values ​​of each fill method according to the fill weight. The missing values ​​in the dimensionality-reduced data are filled with the actual filling values ​​to obtain the filled data.

Images